Artificial Intelligence Nanodegree

Convolutional Neural Networks

Project: Write an Algorithm for a Dog Identification App

rubric


In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!

Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the iPython Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.

In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.

Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.

The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this IPython notebook.


Why We're Here

In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).

Sample Dog Output

In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!

The Road Ahead

We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.

  • Step 0: Import Datasets
  • Step 1: Detect Humans
  • Step 2: Detect Dogs
  • Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
  • Step 4: Use a CNN to Classify Dog Breeds (using Transfer Learning)
  • Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)
  • Step 6: Write your Algorithm
  • Step 7: Test Your Algorithm

Step 0: Import Datasets

Import Dog Dataset

In the code cell below, we import a dataset of dog images. We populate a few variables through the use of the load_files function from the scikit-learn library:

  • train_files, valid_files, test_files - numpy arrays containing file paths to images
  • train_targets, valid_targets, test_targets - numpy arrays containing onehot-encoded classification labels
  • dog_names - list of string-valued dog breed names for translating labels
In [3]:
from sklearn.datasets import load_files       
from keras.utils import np_utils
import numpy as np
from glob import glob
import pandas as pd

# define function to load train, test, and validation datasets
def load_dataset(path):
    data = load_files(path, load_content=False)
    dog_files = np.array(data['filenames'])
    dog_targets = np_utils.to_categorical(np.array(data['target']), 133)
    return dog_files, dog_targets

# load train, test, and validation datasets
train_files, train_targets = load_dataset('dogImages/train')
valid_files, valid_targets = load_dataset('dogImages/valid')
test_files, test_targets = load_dataset('dogImages/test')

# load list of dog names
dog_names = [item[20:-1] for item in sorted(glob("dogImages/train/*/"))]

# print statistics about the dataset
print('There are %d total dog categories.' % len(dog_names))
print('There are %s total dog images.\n' % len(np.hstack([train_files, valid_files, test_files])))
print('There are %d training dog images.' % len(train_files))
print('There are %d validation dog images.' % len(valid_files))
print('There are %d test dog images.'% len(test_files))
Using TensorFlow backend.
There are 133 total dog categories.
There are 8351 total dog images.

There are 6680 training dog images.
There are 835 validation dog images.
There are 836 test dog images.

Import Human Dataset

In the code cell below, we import a dataset of human images, where the file paths are stored in the numpy array human_files.

In [4]:
import random
random.seed(8675309)

# load filenames in shuffled human dataset
human_files = np.array(glob("lfw/*/*"))
random.shuffle(human_files)

# print statistics about the dataset
print('There are %d total human images.' % len(human_files))
There are 13233 total human images.

Step 1: Detect Humans

We use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images. OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory.

In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.

In [5]:
import cv2                
import matplotlib.pyplot as plt                        
%matplotlib inline                               

# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')

# load color (BGR) image
img = cv2.imread(human_files[3])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# find faces in image
faces = face_cascade.detectMultiScale(gray)

# print number of faces detected in the image
print('Number of faces detected:', len(faces))

# get bounding box for each detected face
for (x,y,w,h) in faces:
    # add bounding box to color image
    cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
    
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()
('Number of faces detected:', 1)

Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.

In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.

Write a Human Face Detector

We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.

In [6]:
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
    img = cv2.imread(img_path)
    resized = cv2.resize(img, (224, 224))
    gray = cv2.cvtColor(resized, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray)
    return len(faces) > 0

(IMPLEMENTATION) Assess the Human Face Detector

Question 1: Use the code cell below to test the performance of the face_detector function.

  • What percentage of the first 100 images in human_files have a detected human face?
  • What percentage of the first 100 images in dog_files have a detected human face?

Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.

Answer:

  • 99% of human_files_short images are detected as having at least one human face.
  • 4% of dog_files_short images are detected as having at least one human face.
In [7]:
human_files_short = human_files[:100]
dog_files_short = train_files[:100]
# Do NOT modify the code above this line.

human_files_incorrect = [path for path in human_files_short if not face_detector(path)]
dog_files_incorrect = [path for path in dog_files_short if face_detector(path)]
In [6]:
def show_bad_images(paths):
    fig = plt.figure(figsize=(8, 6))
    for i, path in enumerate(paths):
        a=fig.add_subplot(3,4,i+1)
        plt.imshow(plt.imread(path))
        plt.axis('off')
    plt.show()

print "{} human faces found in human_files_short. incorrect images:".format(100 - len(human_files_incorrect))
show_bad_images(human_files_incorrect)

print "{} human faces found in dog_files_short. incorrect images:".format(len(dog_files_incorrect))
show_bad_images(dog_files_incorrect)
99 human faces found in human_files_short. incorrect images:
4 human faces found in dog_files_short. incorrect images:

Question 2: This algorithmic choice necessitates that we communicate to the user that we accept human images only when they provide a clear view of a face (otherwise, we risk having unneccessarily frustrated users!). In your opinion, is this a reasonable expectation to pose on the user? If not, can you think of a way to detect humans in images that does not necessitate an image with a clearly presented face?

Answer: Requiring a clear view of the face seems like a reasonable requirement for this application.

  1. Most people have some experience with facial recognition applications at this point, both online and in their cameras, so they already have some intuition for software not recognizing faces if they are not clearly visible, at extreme angles or partially obscured.
  2. Since the purpose of the application, when looking at a human face, is to report the dog breed that the face most closely matches, it makes sense to require a clear view of the face. A poor view of the face would make it too difficult to accurately categorize it with a dog breed. This would be an easy requirement to explain to users.

We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on each of the datasets.

Answer: I didn't try any other approaches, but further along in the assignment I was running into issues with python crashing on very large images from my camera. I traced it back and realized that we were running the cv2 cascade classifier on the original image without any scaling, so I added a step to scale the image to (224, 224) in the face_detector function. Not only did this fix the crashing, it also caused a nice increase in accuracy, going from 98% to 99% on true positives for humans and 11% to 4% on false positives for dogs. The classifier seems to be highly sensitive to the size of human/dog faces in the input images.


Step 2: Detect Dogs

In this section, we use a pre-trained ResNet-50 model to detect dogs in images. Our first line of code downloads the ResNet-50 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories. Given an image, this pre-trained ResNet-50 model returns a prediction (derived from the available categories in ImageNet) for the object that is contained in the image.

In [8]:
from keras.applications.resnet50 import ResNet50

# define ResNet50 model
ResNet50_model = ResNet50(weights='imagenet')
Downloading data from https://github.com/fchollet/deep-learning-models/releases/download/v0.2/resnet50_weights_tf_dim_ordering_tf_kernels.h5

Pre-process the Data

When using TensorFlow as backend, Keras CNNs require a 4D array (which we'll also refer to as a 4D tensor) as input, with shape

$$ (\text{nb_samples}, \text{rows}, \text{columns}, \text{channels}), $$

where nb_samples corresponds to the total number of images (or samples), and rows, columns, and channels correspond to the number of rows, columns, and channels for each image, respectively.

The path_to_tensor function below takes a string-valued file path to a color image as input and returns a 4D tensor suitable for supplying to a Keras CNN. The function first loads the image and resizes it to a square image that is $224 \times 224$ pixels. Next, the image is converted to an array, which is then resized to a 4D tensor. In this case, since we are working with color images, each image has three channels. Likewise, since we are processing a single image (or sample), the returned tensor will always have shape

$$ (1, 224, 224, 3). $$

The paths_to_tensor function takes a numpy array of string-valued image paths as input and returns a 4D tensor with shape

$$ (\text{nb_samples}, 224, 224, 3). $$

Here, nb_samples is the number of samples, or number of images, in the supplied array of image paths. It is best to think of nb_samples as the number of 3D tensors (where each 3D tensor corresponds to a different image) in your dataset!

In [9]:
from keras.preprocessing import image                  
from tqdm import tqdm

def path_to_tensor(img_path):
    # loads RGB image as PIL.Image.Image type
    img = image.load_img(img_path, target_size=(224, 224))
    # convert PIL.Image.Image type to 3D tensor with shape (224, 224, 3)
    x = image.img_to_array(img)
    # convert 3D tensor to 4D tensor with shape (1, 224, 224, 3) and return 4D tensor
    return np.expand_dims(x, axis=0)

def paths_to_tensor(img_paths):
    list_of_tensors = [path_to_tensor(img_path) for img_path in tqdm(img_paths)]
    return np.vstack(list_of_tensors)

Making Predictions with ResNet-50

Getting the 4D tensor ready for ResNet-50, and for any other pre-trained model in Keras, requires some additional processing. First, the RGB image is converted to BGR by reordering the channels. All pre-trained models have the additional normalization step that the mean pixel (expressed in RGB as $[103.939, 116.779, 123.68]$ and calculated from all pixels in all images in ImageNet) must be subtracted from every pixel in each image. This is implemented in the imported function preprocess_input. If you're curious, you can check the code for preprocess_input here.

Now that we have a way to format our image for supplying to ResNet-50, we are now ready to use the model to extract the predictions. This is accomplished with the predict method, which returns an array whose $i$-th entry is the model's predicted probability that the image belongs to the $i$-th ImageNet category. This is implemented in the ResNet50_predict_labels function below.

By taking the argmax of the predicted probability vector, we obtain an integer corresponding to the model's predicted object class, which we can identify with an object category through the use of this dictionary.

In [10]:
from keras.applications.resnet50 import preprocess_input, decode_predictions

def ResNet50_predict_labels(img_path):
    # returns prediction vector for image located at img_path
    img = preprocess_input(path_to_tensor(img_path))
    return np.argmax(ResNet50_model.predict(img))

Write a Dog Detector

While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained ResNet-50 model, we need only check if the ResNet50_predict_labels function above returns a value between 151 and 268 (inclusive).

We use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).

In [11]:
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
    prediction = ResNet50_predict_labels(img_path)
    return ((prediction <= 268) & (prediction >= 151)) 

(IMPLEMENTATION) Assess the Dog Detector

Question 3: Use the code cell below to test the performance of your dog_detector function.

  • What percentage of the images in human_files_short have a detected dog?
  • What percentage of the images in dog_files_short have a detected dog?

Answer:

  • 1% of human_files_short images are detected as having a dog face.
  • 100% of dog_files_short images are detected as having a dog face.
In [12]:
human_files_incorrect_dogs = [path for path in human_files_short if dog_detector(path)]
dog_files_incorrect_dogs = [path for path in dog_files_short if not dog_detector(path)]
In [12]:
print "{} dog faces found in human_files_short. incorrect images:".format(len(human_files_incorrect_dogs))
show_bad_images(human_files_incorrect_dogs)

print "{} dog faces found in dog_files_short. incorrect images:".format(100 - len(dog_files_incorrect_dogs))
show_bad_images(dog_files_incorrect_dogs)
1 dog faces found in human_files_short. incorrect images:
100 dog faces found in dog_files_short. incorrect images:
<matplotlib.figure.Figure at 0x7f4914720990>

Step 3: Create a CNN to Classify Dog Breeds (from Scratch)

Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 1%. In Step 5 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.

Be careful with adding too many trainable layers! More parameters means longer training, which means you are more likely to need a GPU to accelerate the training process. Thankfully, Keras provides a handy estimate of the time that each epoch is likely to take; you can extrapolate this estimate to figure out how long it will take for your algorithm to train.

We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have great difficulty in distinguishing between a Brittany and a Welsh Springer Spaniel.

Brittany Welsh Springer Spaniel

It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).

Curly-Coated Retriever American Water Spaniel

Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.

Yellow Labrador Chocolate Labrador Black Labrador

We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.

Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!

Pre-process the Data

We rescale the images by dividing every pixel in every image by 255.

In [13]:
from PIL import ImageFile                            
ImageFile.LOAD_TRUNCATED_IMAGES = True                 

# pre-process the data for Keras
train_tensors = paths_to_tensor(train_files).astype('float32')/255
valid_tensors = paths_to_tensor(valid_files).astype('float32')/255
test_tensors = paths_to_tensor(test_files).astype('float32')/255
100%|██████████| 6680/6680 [01:02<00:00, 107.32it/s]
100%|██████████| 835/835 [00:08<00:00, 98.51it/s] 
100%|██████████| 836/836 [00:07<00:00, 104.67it/s]

(IMPLEMENTATION) Model Architecture

Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:

    model.summary()

We have imported some Python modules to get you started, but feel free to import as many modules as you need. If you end up getting stuck, here's a hint that specifies a model that trains relatively fast on CPU and attains >1% test accuracy in 5 epochs:

Sample CNN

Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. If you chose to use the hinted architecture above, describe why you think that CNN architecture should work well for the image classification task.

Answer: My first attempt reached 2% accuracy. I based my choices for the model on a combination of the "CNNs in Keras: Practical Example" video in the course, along with the hint above.

The architecture is a series of three Conv2D/MaxPooling2D pairs. Each Conv2D uses 'relu' activation, as that was highly recommended for CNNs in the course lectures. I enabled padding='same' as well, since it was mentioned in the lectures that it tends to perform a little better than the default of 'valid'. Each MaxPooling layer downscales the input by 1/2 in each direction, resulting in 1/4th of the size after each pool.

The first stage is a 16-filter Conv2D layer with the input image shape. Subsequent states are 32- and then 64-filter layers, following the theory that we want to start with shallower, more spatial layers and work our way towards much deeper but narrower layers that have gotten rid of the spatial information in favor of higher-level information.

Following these pairs of layers I added a Global Average Pooling layer to take the average of all the filter positions from the last layer, which should help with identifying features regardless of their position.

I then added a 250-node dense layer, to learn from the features discovered in the previous layers. This feeds into a smaller Dense output layer with the same number of nodes as the number of categories, with softmax activation, to do the final probability-based categorization.

While I was doing some further reading I discovered the recent use of Batch Normalization in CNNs after each convolution step, in order to normalize values in a deep NN to avoid very small or large values. This allows for faster training and thus ultimately higher accuracy from the same training. I applied a default BatchNormalization layer after each Conv2D layer, and the accuracy improved to 9%. I read up on some other potential improvements as well but decided to move on for now, since the target was only 1% and this is a slow process on my CPU.

In [14]:
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential
from keras.layers.normalization import BatchNormalization

model = Sequential()

### TODO: Define your architecture.
model.add(Conv2D(16, 2, activation='relu', padding='same', input_shape=(224, 224, 3)))
model.add(MaxPooling2D(pool_size=2))
model.add(BatchNormalization())
model.add(Conv2D(32, 2, activation='relu', padding='same'))
model.add(MaxPooling2D(pool_size=2))
model.add(BatchNormalization())
model.add(Conv2D(64, 2, activation='relu', padding='same'))
model.add(MaxPooling2D(pool_size=2))
model.add(BatchNormalization())
model.add(GlobalAveragePooling2D())
model.add(Dense(250, activation='relu'))
model.add(Dense(len(dog_names), activation='softmax'))

model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 224, 224, 16)      208       
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 112, 112, 16)      0         
_________________________________________________________________
batch_normalization_1 (Batch (None, 112, 112, 16)      64        
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 112, 112, 32)      2080      
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 56, 56, 32)        0         
_________________________________________________________________
batch_normalization_2 (Batch (None, 56, 56, 32)        128       
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 56, 56, 64)        8256      
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 28, 28, 64)        0         
_________________________________________________________________
batch_normalization_3 (Batch (None, 28, 28, 64)        256       
_________________________________________________________________
global_average_pooling2d_1 ( (None, 64)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 250)               16250     
_________________________________________________________________
dense_2 (Dense)              (None, 133)               33383     
=================================================================
Total params: 60,625.0
Trainable params: 60,401.0
Non-trainable params: 224.0
_________________________________________________________________

Compile the Model

In [15]:
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

(IMPLEMENTATION) Train the Model

Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.

You are welcome to augment the training data, but this is not a requirement.

In [16]:
from keras.callbacks import ModelCheckpoint  

### specify the number of epochs that you would like to use to train the model.

epochs = 10

### Do NOT modify the code below this line.

checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.from_scratch.hdf5', 
                               verbose=1, save_best_only=True)

model.fit(train_tensors, train_targets, 
          validation_data=(valid_tensors, valid_targets),
          epochs=epochs, batch_size=20, callbacks=[checkpointer], verbose=1)
Train on 6680 samples, validate on 835 samples
Epoch 1/10
6660/6680 [============================>.] - ETA: 0s - loss: 4.7848 - acc: 0.0219      Epoch 00000: val_loss improved from inf to 6.68550, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 261s - loss: 4.7848 - acc: 0.0219 - val_loss: 6.6855 - val_acc: 0.0096
Epoch 2/10
6660/6680 [============================>.] - ETA: 0s - loss: 4.5316 - acc: 0.0386      Epoch 00001: val_loss improved from 6.68550 to 4.99173, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 244s - loss: 4.5308 - acc: 0.0385 - val_loss: 4.9917 - val_acc: 0.0180
Epoch 3/10
6660/6680 [============================>.] - ETA: 0s - loss: 4.3817 - acc: 0.0536      Epoch 00002: val_loss improved from 4.99173 to 4.73252, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 247s - loss: 4.3817 - acc: 0.0537 - val_loss: 4.7325 - val_acc: 0.0347
Epoch 4/10
6660/6680 [============================>.] - ETA: 0s - loss: 4.2633 - acc: 0.0707      Epoch 00003: val_loss improved from 4.73252 to 4.59102, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 256s - loss: 4.2639 - acc: 0.0705 - val_loss: 4.5910 - val_acc: 0.0479
Epoch 5/10
6660/6680 [============================>.] - ETA: 0s - loss: 4.1576 - acc: 0.0763  Epoch 00004: val_loss improved from 4.59102 to 4.46697, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 244s - loss: 4.1590 - acc: 0.0763 - val_loss: 4.4670 - val_acc: 0.0563
Epoch 6/10
6660/6680 [============================>.] - ETA: 0s - loss: 4.0470 - acc: 0.0911  Epoch 00005: val_loss improved from 4.46697 to 4.31226, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 244s - loss: 4.0475 - acc: 0.0913 - val_loss: 4.3123 - val_acc: 0.0766
Epoch 7/10
6660/6680 [============================>.] - ETA: 0s - loss: 3.9462 - acc: 0.1075  Epoch 00006: val_loss did not improve
6680/6680 [==============================] - 243s - loss: 3.9471 - acc: 0.1073 - val_loss: 4.5322 - val_acc: 0.0707
Epoch 8/10
6660/6680 [============================>.] - ETA: 0s - loss: 3.8415 - acc: 0.1167  Epoch 00007: val_loss did not improve
6680/6680 [==============================] - 244s - loss: 3.8403 - acc: 0.1165 - val_loss: 4.3885 - val_acc: 0.0910
Epoch 9/10
6660/6680 [============================>.] - ETA: 0s - loss: 3.7522 - acc: 0.1341  Epoch 00008: val_loss improved from 4.31226 to 4.18774, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 230s - loss: 3.7531 - acc: 0.1341 - val_loss: 4.1877 - val_acc: 0.0898
Epoch 10/10
6660/6680 [============================>.] - ETA: 0s - loss: 3.6474 - acc: 0.1464  Epoch 00009: val_loss improved from 4.18774 to 4.18710, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 232s - loss: 3.6475 - acc: 0.1461 - val_loss: 4.1871 - val_acc: 0.1030
Out[16]:
<keras.callbacks.History at 0x7f48e2c5f510>

Load the Model with the Best Validation Loss

In [17]:
model.load_weights('saved_models/weights.best.from_scratch.hdf5')

Test the Model

Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 1%.

In [18]:
# get index of predicted dog breed for each image in test set
dog_breed_predictions = [np.argmax(model.predict(np.expand_dims(tensor, axis=0))) for tensor in test_tensors]

# report test accuracy
test_accuracy = 100*np.sum(np.array(dog_breed_predictions)==np.argmax(test_targets, axis=1))/len(dog_breed_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
Test accuracy: 9.0000%

Step 4: Use a CNN to Classify Dog Breeds

To reduce training time without sacrificing accuracy, we show you how to train a CNN using transfer learning. In the following step, you will get a chance to use transfer learning to train your own CNN.

Obtain Bottleneck Features

In [19]:
bottleneck_features = np.load('bottleneck_features/DogVGG16Data.npz')
train_VGG16 = bottleneck_features['train']
valid_VGG16 = bottleneck_features['valid']
test_VGG16 = bottleneck_features['test']

Model Architecture

The model uses the the pre-trained VGG-16 model as a fixed feature extractor, where the last convolutional output of VGG-16 is fed as input to our model. We only add a global average pooling layer and a fully connected layer, where the latter contains one node for each dog category and is equipped with a softmax.

In [20]:
VGG16_model = Sequential()
VGG16_model.add(GlobalAveragePooling2D(input_shape=train_VGG16.shape[1:]))
VGG16_model.add(Dense(133, activation='softmax'))

VGG16_model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
global_average_pooling2d_2 ( (None, 512)               0         
_________________________________________________________________
dense_3 (Dense)              (None, 133)               68229     
=================================================================
Total params: 68,229.0
Trainable params: 68,229.0
Non-trainable params: 0.0
_________________________________________________________________

Compile the Model

In [21]:
VGG16_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

Train the Model

In [22]:
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.VGG16.hdf5', 
                               verbose=1, save_best_only=True)

VGG16_model.fit(train_VGG16, train_targets, 
          validation_data=(valid_VGG16, valid_targets),
          epochs=20, batch_size=20, callbacks=[checkpointer], verbose=1)
Train on 6680 samples, validate on 835 samples
Epoch 1/20
6400/6680 [===========================>..] - ETA: 0s - loss: 12.0557 - acc: 0.1248 Epoch 00000: val_loss improved from inf to 10.26627, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 12.0028 - acc: 0.1281 - val_loss: 10.2663 - val_acc: 0.2299
Epoch 2/20
6480/6680 [============================>.] - ETA: 0s - loss: 9.6628 - acc: 0.3046 Epoch 00001: val_loss improved from 10.26627 to 9.63467, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 9.6532 - acc: 0.3063 - val_loss: 9.6347 - val_acc: 0.2970
Epoch 3/20
6560/6680 [============================>.] - ETA: 0s - loss: 9.0920 - acc: 0.3726 Epoch 00002: val_loss improved from 9.63467 to 9.27225, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 9.1110 - acc: 0.3716 - val_loss: 9.2723 - val_acc: 0.3257
Epoch 4/20
6520/6680 [============================>.] - ETA: 0s - loss: 8.8343 - acc: 0.4018Epoch 00003: val_loss improved from 9.27225 to 9.11207, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.8289 - acc: 0.4021 - val_loss: 9.1121 - val_acc: 0.3473
Epoch 5/20
6520/6680 [============================>.] - ETA: 0s - loss: 8.6890 - acc: 0.4271Epoch 00004: val_loss improved from 9.11207 to 9.05713, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.6936 - acc: 0.4266 - val_loss: 9.0571 - val_acc: 0.3533
Epoch 6/20
6480/6680 [============================>.] - ETA: 0s - loss: 8.6219 - acc: 0.4383 Epoch 00005: val_loss improved from 9.05713 to 9.03954, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.6325 - acc: 0.4373 - val_loss: 9.0395 - val_acc: 0.3665
Epoch 7/20
6400/6680 [===========================>..] - ETA: 0s - loss: 8.5662 - acc: 0.4480Epoch 00006: val_loss improved from 9.03954 to 8.95819, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.5907 - acc: 0.4466 - val_loss: 8.9582 - val_acc: 0.3880
Epoch 8/20
6480/6680 [============================>.] - ETA: 0s - loss: 8.5514 - acc: 0.4540Epoch 00007: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 8.5625 - acc: 0.4534 - val_loss: 8.9985 - val_acc: 0.3880
Epoch 9/20
6480/6680 [============================>.] - ETA: 0s - loss: 8.4858 - acc: 0.4580Epoch 00008: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 8.4793 - acc: 0.4585 - val_loss: 9.0508 - val_acc: 0.3653
Epoch 10/20
6620/6680 [============================>.] - ETA: 0s - loss: 8.3969 - acc: 0.4633Epoch 00009: val_loss improved from 8.95819 to 8.93875, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 3s - loss: 8.4116 - acc: 0.4624 - val_loss: 8.9387 - val_acc: 0.3737
Epoch 11/20
6480/6680 [============================>.] - ETA: 0s - loss: 8.3515 - acc: 0.4691 Epoch 00010: val_loss improved from 8.93875 to 8.90739, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.3806 - acc: 0.4675 - val_loss: 8.9074 - val_acc: 0.3856
Epoch 12/20
6420/6680 [===========================>..] - ETA: 0s - loss: 8.3028 - acc: 0.4737Epoch 00011: val_loss improved from 8.90739 to 8.73666, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.3200 - acc: 0.4725 - val_loss: 8.7367 - val_acc: 0.4060
Epoch 13/20
6480/6680 [============================>.] - ETA: 0s - loss: 8.2458 - acc: 0.4804Epoch 00012: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 8.2391 - acc: 0.4808 - val_loss: 8.8400 - val_acc: 0.3928
Epoch 14/20
6600/6680 [============================>.] - ETA: 0s - loss: 8.2265 - acc: 0.4818 Epoch 00013: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 8.2317 - acc: 0.4816 - val_loss: 8.7481 - val_acc: 0.4108
Epoch 15/20
6600/6680 [============================>.] - ETA: 0s - loss: 8.2171 - acc: 0.4847 Epoch 00014: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 8.2201 - acc: 0.4846 - val_loss: 8.7401 - val_acc: 0.4000
Epoch 16/20
6540/6680 [============================>.] - ETA: 0s - loss: 8.2172 - acc: 0.4862Epoch 00015: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 8.2141 - acc: 0.4864 - val_loss: 8.7432 - val_acc: 0.4000
Epoch 17/20
6560/6680 [============================>.] - ETA: 0s - loss: 8.2099 - acc: 0.4863Epoch 00016: val_loss did not improve
6680/6680 [==============================] - 4s - loss: 8.1908 - acc: 0.4874 - val_loss: 8.7566 - val_acc: 0.4000
Epoch 18/20
6620/6680 [============================>.] - ETA: 0s - loss: 8.1335 - acc: 0.4893Epoch 00017: val_loss improved from 8.73666 to 8.66124, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.1328 - acc: 0.4894 - val_loss: 8.6612 - val_acc: 0.4024
Epoch 19/20
6580/6680 [============================>.] - ETA: 0s - loss: 8.0422 - acc: 0.4907Epoch 00018: val_loss improved from 8.66124 to 8.55582, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 2s - loss: 8.0473 - acc: 0.4906 - val_loss: 8.5558 - val_acc: 0.4000
Epoch 20/20
6520/6680 [============================>.] - ETA: 0s - loss: 7.8504 - acc: 0.5005Epoch 00019: val_loss improved from 8.55582 to 8.43937, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 2s - loss: 7.8616 - acc: 0.4997 - val_loss: 8.4394 - val_acc: 0.4060
Out[22]:
<keras.callbacks.History at 0x7f4914705bd0>

Load the Model with the Best Validation Loss

In [23]:
VGG16_model.load_weights('saved_models/weights.best.VGG16.hdf5')

Test the Model

Now, we can use the CNN to test how well it identifies breed within our test dataset of dog images. We print the test accuracy below.

In [24]:
# get index of predicted dog breed for each image in test set
VGG16_predictions = [np.argmax(VGG16_model.predict(np.expand_dims(feature, axis=0))) for feature in test_VGG16]

# report test accuracy
test_accuracy = 100*np.sum(np.array(VGG16_predictions)==np.argmax(test_targets, axis=1))/len(VGG16_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
Test accuracy: 41.0000%

Predict Dog Breed with the Model

In [25]:
from extract_bottleneck_features import *

def VGG16_predict_breed(img_path):
    # extract bottleneck features
    bottleneck_feature = extract_VGG16(path_to_tensor(img_path))
    # obtain predicted vector
    predicted_vector = VGG16_model.predict(bottleneck_feature)
    # return dog breed that is predicted by the model
    return dog_names[np.argmax(predicted_vector)]

Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)

You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.

In Step 4, we used transfer learning to create a CNN using VGG-16 bottleneck features. In this section, you must use the bottleneck features from a different pre-trained model. To make things easier for you, we have pre-computed the features for all of the networks that are currently available in Keras:

The files are encoded as such:

Dog{network}Data.npz

where {network}, in the above filename, can be one of VGG19, Resnet50, InceptionV3, or Xception. Pick one of the above architectures, download the corresponding bottleneck features, and store the downloaded file in the bottleneck_features/ folder in the repository.

(IMPLEMENTATION) Obtain Bottleneck Features

In the code block below, extract the bottleneck features corresponding to the train, test, and validation sets by running the following:

bottleneck_features = np.load('bottleneck_features/Dog{network}Data.npz')
train_{network} = bottleneck_features['train']
valid_{network} = bottleneck_features['valid']
test_{network} = bottleneck_features['test']
In [13]:
bottleneck_features = np.load('bottleneck_features/DogResnet50Data.npz')
train_Resnet50 = bottleneck_features['train']
valid_Resnet50 = bottleneck_features['valid']
test_Resnet50 = bottleneck_features['test']

(IMPLEMENTATION) Model Architecture

Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:

    <your model's name>.summary()

Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.

Answer:

I based the architecture on information in the "Transfer Learning" section of the course lectures. The training data set is small-ish, at 6680 samples, so I followed the advice from "Small Data Set, Similar Data". It also makes sense to do it this way simply because I don't have the full ability to re-train the earlier layers, as outlined in "Large Data Set, Similar Data".

So, I first added a global average pool, similar to my from-scratch network, and then a Dense layer with the same number of nodes as the number of categories, with softmax activation, to do the final probability-based categorization.

I initially tried adding another dense layer before the final one, with 200-250 nodes, but this did not increase accuracy.

Switching from "rmsprop" to "adam" for the optimizer boosted accuracy from 80% to 82%.

Suitability

My final architecture uses transfer learning with the majority of the network coming from ResNet [1], a highly successful CNN from Microsoft that won the ImageNet competition in 2015 and outperforms humans on the ImageNet data set. I am using the shallower 50-layer version of ResNet. The ImageNet data already contains dog classifications, so my problem is of a very similar nature. This means that an already-trained ResNet CNN is highly suitable to my problem, I just need to adjust the final layers of the network to output the classifications that I am looking for, taking advantage of the features the ResNet has already learned.

Because of this I start with already a very good base network for identifying objects in images of this type. Then by adding just the tail end of my initial from-scratch network, I am able to utilize ResNet50's already-high performance for my specific dog classification problem.

[1] https://arxiv.org/abs/1512.03385

In [16]:
my_Resnet50_model = Sequential()
my_Resnet50_model.add(GlobalAveragePooling2D(input_shape=train_Resnet50.shape[1:]))
my_Resnet50_model.add(Dense(len(dog_names), activation='softmax'))

my_Resnet50_model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
global_average_pooling2d_2 ( (None, 2048)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 133)               272517    
=================================================================
Total params: 272,517.0
Trainable params: 272,517.0
Non-trainable params: 0.0
_________________________________________________________________

(IMPLEMENTATION) Compile the Model

In [17]:
my_Resnet50_model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

(IMPLEMENTATION) Train the Model

Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.

You are welcome to augment the training data, but this is not a requirement.

In [21]:
checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.Resnet50.hdf5', 
                               verbose=1, save_best_only=True)

my_Resnet50_model.fit(train_Resnet50, train_targets, 
          validation_data=(valid_Resnet50, valid_targets),
          epochs=15, batch_size=20, callbacks=[checkpointer], verbose=1)
Train on 6680 samples, validate on 835 samples
Epoch 1/15
6520/6680 [============================>.] - ETA: 0s - loss: 0.0750 - acc: 0.9836 Epoch 00000: val_loss improved from inf to 0.72089, saving model to saved_models/weights.best.Resnet50.hdf5
6680/6680 [==============================] - 1s - loss: 0.0751 - acc: 0.9837 - val_loss: 0.7209 - val_acc: 0.8240
Epoch 2/15
6640/6680 [============================>.] - ETA: 0s - loss: 0.0352 - acc: 0.9932Epoch 00001: val_loss improved from 0.72089 to 0.63893, saving model to saved_models/weights.best.Resnet50.hdf5
6680/6680 [==============================] - 1s - loss: 0.0351 - acc: 0.9933 - val_loss: 0.6389 - val_acc: 0.8323
Epoch 3/15
6520/6680 [============================>.] - ETA: 0s - loss: 0.0227 - acc: 0.9957Epoch 00002: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0224 - acc: 0.9958 - val_loss: 0.7121 - val_acc: 0.8012
Epoch 4/15
6500/6680 [============================>.] - ETA: 0s - loss: 0.0191 - acc: 0.9966Epoch 00003: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0195 - acc: 0.9964 - val_loss: 0.6627 - val_acc: 0.8192
Epoch 5/15
6480/6680 [============================>.] - ETA: 0s - loss: 0.0212 - acc: 0.9952Epoch 00004: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0213 - acc: 0.9951 - val_loss: 0.7032 - val_acc: 0.8120
Epoch 6/15
6600/6680 [============================>.] - ETA: 0s - loss: 0.0183 - acc: 0.9964Epoch 00005: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0183 - acc: 0.9963 - val_loss: 0.7155 - val_acc: 0.8192
Epoch 7/15
6460/6680 [============================>.] - ETA: 0s - loss: 0.0311 - acc: 0.9912Epoch 00006: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0325 - acc: 0.9912 - val_loss: 0.8446 - val_acc: 0.8024
Epoch 8/15
6440/6680 [===========================>..] - ETA: 0s - loss: 0.0591 - acc: 0.9856Epoch 00007: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0596 - acc: 0.9853 - val_loss: 1.0261 - val_acc: 0.7701
Epoch 9/15
6660/6680 [============================>.] - ETA: 0s - loss: 0.0420 - acc: 0.9884Epoch 00008: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0420 - acc: 0.9883 - val_loss: 0.8338 - val_acc: 0.8132
Epoch 10/15
6560/6680 [============================>.] - ETA: 0s - loss: 0.0109 - acc: 0.9977Epoch 00009: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0108 - acc: 0.9978 - val_loss: 0.8217 - val_acc: 0.8204
Epoch 11/15
6600/6680 [============================>.] - ETA: 0s - loss: 0.0115 - acc: 0.9974    Epoch 00010: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0118 - acc: 0.9972 - val_loss: 0.7435 - val_acc: 0.8371
Epoch 12/15
6560/6680 [============================>.] - ETA: 0s - loss: 0.0158 - acc: 0.9957    Epoch 00011: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0157 - acc: 0.9958 - val_loss: 0.8029 - val_acc: 0.8108
Epoch 13/15
6600/6680 [============================>.] - ETA: 0s - loss: 0.0441 - acc: 0.9882    Epoch 00012: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0447 - acc: 0.9882 - val_loss: 0.9701 - val_acc: 0.7952
Epoch 14/15
6560/6680 [============================>.] - ETA: 0s - loss: 0.0207 - acc: 0.9944Epoch 00013: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0204 - acc: 0.9945 - val_loss: 0.8546 - val_acc: 0.8144
Epoch 15/15
6460/6680 [============================>.] - ETA: 0s - loss: 0.0153 - acc: 0.9963Epoch 00014: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 0.0156 - acc: 0.9963 - val_loss: 0.8072 - val_acc: 0.8156
Out[21]:
<keras.callbacks.History at 0x7f862295da50>

(IMPLEMENTATION) Load the Model with the Best Validation Loss

In [22]:
my_Resnet50_model.load_weights('saved_models/weights.best.Resnet50.hdf5')

(IMPLEMENTATION) Test the Model

Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 60%.

In [23]:
# get index of predicted dog breed for each image in test set
Resnet50_predictions = [np.argmax(my_Resnet50_model.predict(np.expand_dims(feature, axis=0))) for feature in test_Resnet50]

# report test accuracy
test_accuracy = 100*np.sum(np.array(Resnet50_predictions)==np.argmax(test_targets, axis=1))/len(Resnet50_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
Test accuracy: 82.0000%

(IMPLEMENTATION) Predict Dog Breed with the Model

Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan_hound, etc) that is predicted by your model.

Similar to the analogous function in Step 5, your function should have three steps:

  1. Extract the bottleneck features corresponding to the chosen CNN model.
  2. Supply the bottleneck features as input to the model to return the predicted vector. Note that the argmax of this prediction vector gives the index of the predicted dog breed.
  3. Use the dog_names array defined in Step 0 of this notebook to return the corresponding breed.

The functions to extract the bottleneck features can be found in extract_bottleneck_features.py, and they have been imported in an earlier code cell. To obtain the bottleneck features corresponding to your chosen CNN architecture, you need to use the function

extract_{network}

where {network}, in the above filename, should be one of VGG19, Resnet50, InceptionV3, or Xception.

In [26]:
from extract_bottleneck_features import *

def Resnet50_predict_breed(img_path):
    # extract bottleneck features
    bottleneck_feature = extract_Resnet50(path_to_tensor(img_path))
    # obtain predicted vector
    predicted_vector = my_Resnet50_model.predict(bottleneck_feature)
    # return dog breed that is predicted by the model
    return dog_names[np.argmax(predicted_vector)]
In [43]:
# Just testing things out

small_test_indices = np.random.choice(range(len(test_files)), 8)
small_test_paths = test_files[small_test_indices]
small_test_predictions = map(Resnet50_predict_breed, small_test_paths)
small_test_labels = [dog_names[np.argmax(lv)] for lv in test_targets[small_test_indices]]
In [44]:
print("Resnet50 predictions and labels")
plt.figure(figsize=(10, 6))
plt.subplots_adjust(bottom=0, left=.01, right=.99, top=.90, hspace=.35)
for i, path in enumerate(small_test_paths):
    plt.subplot(2,4,i+1)
    plt.imshow(plt.imread(path))
    plt.title("pred: {}\nact: {}".format(small_test_predictions[i], small_test_labels[i]))
    plt.axis('off')
plt.show()
Resnet50 predictions and labels

Step 6: Write your Algorithm

Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,

  • if a dog is detected in the image, return the predicted breed.
  • if a human is detected in the image, return the resembling dog breed.
  • if neither is detected in the image, provide output that indicates an error.

You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and dog_detector functions developed above. You are required to use your CNN from Step 5 to predict dog breed.

Some sample output for our algorithm is provided below, but feel free to design your own user experience!

Sample Human Output

(IMPLEMENTATION) Write your Algorithm

In [35]:
def predict(path):
    plt.figure(figsize=(10, 6))
    plt.imshow(plt.imread(path))
    if dog_detector(path):
        breed = Resnet50_predict_breed(path)
        plt.title("good doggo!\nyou are a...\n{}".format(breed))
    elif face_detector(path):
        breed = Resnet50_predict_breed(path)
        plt.title("hello, human!\nyou look like a...\n{}".format(breed))
    else:
        plt.title("no humans or doggins found in the image. are you sure it's not a cat?")
    plt.axis('off')
    plt.show()

Step 7: Test Your Algorithm

In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?

(IMPLEMENTATION) Test Your Algorithm on Sample Images!

Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.

Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.

Answer:

Overall the algorithm performs pretty well. It feels pretty random which dog breed it returns for people, but of course most people don't really look much like a particular dog breed anyway, so it makes sense. I was impressed that the algorithm is correct when I give it a picture of my cat: neither dog nor human is found. There's a more serious problem with the face detection, discussed below.

Possible improvements:

  1. The OpenCV Haar Detection algorithm for finding faces seems to have a lot of trouble finding faces on dark-skinned people. I've found some discussion of this on the Internet here and here. This is super unfortunate, I'd certainly explore other algorithms for face detection in a commercial product. This Medium post looks like it might be a good start, only step 1 (find faces in the photo) would be needed.
  2. In my earlier testing, the face detection has an 11% false positive rate when given pictures of dogs. This was problematic with my first version of the predict function, because I tested for human faces first, and many of my pictures of dogs were getting tagged as human. I ended up reversing the order to run dog_detector first, since this has a better FP rate. UPDATE: As I discussed earlier, I ended up adding a step to face_detector to resize the image, in order to fix my python VM crashing on very large input images. This also improved FP performance from 11% to 4%.
  3. The algorithm recognizes 133 breeds of dog, but in reality there are many more widely-recognized breeds that this training data did not include. For instance, I tested a few pictures of my dog Sen who is a Catahoula Leopard Dog. This breed wasn't in the data set, so it's not surprising that she is recognized as a different breed in every image (Italian Greyhound, for example). In a commercial product I'd want to expand the training data to include other breeds as well. I am curious if this would cause problems with using transfer learning from ResNet50, since that NN may not have ever learned the features necessary to identify less common breeds very well.

I like the idea proposed in the rubric, of returning multiple breeds with some confidence factor, since the majority of dogs are mutts mixing multiple breeds. I am going to implement that below.

In [36]:
predict("my_images/IMG_1312.jpg")
predict("my_images/IMG_1231.jpg")
predict("my_images/DSC00559.JPG")
predict("my_images/DSC00142.JPG")
predict("my_images/robin.jpg")
predict("my_images/bean.jpg")
predict("my_images/DSC00287.JPG")
predict("my_images/tracy.jpg")
predict("my_images/snoop.jpg")
In [39]:
def Resnet50_predict_breeds(img_path):
    # extract bottleneck features
    bottleneck_feature = extract_Resnet50(path_to_tensor(img_path))
    # obtain predicted vector
    predicted_vector = my_Resnet50_model.predict(bottleneck_feature)
    # return dog breeds that are highly predicted by the model
    max_p = np.max(predicted_vector)
    breeds = [dog_names[np.argmax(predicted_vector)]]
    for idx,p in enumerate(predicted_vector[0]):
        if p / max_p > 0.4 and dog_names[idx] != breeds[0]:
            breeds.append(dog_names[idx])
    return breeds

def predict_mutts(path):
    plt.figure(figsize=(10, 6))
    plt.imshow(plt.imread(path))
    if dog_detector(path):
        breed = Resnet50_predict_breeds(path)
        plt.title("good doggo!\nyou are a...\n{}".format("\n".join(breed)))
    elif face_detector(path):
        breed = Resnet50_predict_breeds(path)
        plt.title("hello, human!\nyou look like a...\n{}".format("\n".join(breed)))
    else:
        plt.title("no humans or doggins found in the image. are you sure it's not a cat?")
    plt.axis('off')
    plt.show()
In [42]:
predict_mutts("my_images/mutt1.jpg")
predict_mutts("my_images/mutt3.jpg")

I attempted to predict mutt breed mixes by finding the highest-scoring predicted breed, as before, but then also adding to the list any other breed that was predicted with at least 40% of the probability of the highest score.

Overall it doesn't seem to have worked very well with this specific algorithm, I'm not too surprised. Visually determining mutt pedigree is notoriously difficult even for people who work with dogs all the time.

In [ ]: